4.3. Lasso for Dimensionality Reduction of Features

Shi et.al [23] proved that the least absolute shrinkage and selection operator (Lasso) method can effectively reduce information redundancy and delete some unimportant features compared with principal components analysis (PCA), ReliefF, and Elastic net. Therefore, we use Lasso as the dimensionality reduction algorithm for this paper. LASSO proposed by Tibshirani [77] is a compression estimation method with l1 regularization implemented to achieve a sparse solution. LASSO is used to perform feature selection by forcing many parameters corresponding to the irrelevant and redundant features to zero value, and retaining the features corresponding to the non-zero coefficients for subsequent classification [78,79,80]. The aim of this approach is to minimize the cost function:

where yn represents the corresponding response vector of a DTI pair, that is, the class label of the sample, N is the number of samples, xnm is the m-th feature of the n-th sample, λ is the regularization parameter, and βm is the regression coefficients of m-th feature [78].

Therefore, through formula (13), we eliminate the noise and redundant information contained in the high-dimensional data obtained after the original drug and target feature extraction